wuu 开源桌面应用发布预览版：让 AI 智能体在本地项目中工作

An open-source desktop app where AI agents work on your local projects,
and every file they read, command they run and line they change stays in view.

Download for Mac · Documentation · 简体中文

Point wuu at a folder, describe the task, and let an agent read the code, run commands and edit files. The conversation shows each step as it happens. Beside it, Review shows the actual diff on disk, so you check the result rather than taking the summary on trust.

See it in action

A user reports that October 1 expenses land in September. The agent searches the code, runs the tests to reproduce the bug, fixes the month calculation and adds a test. One click opens the change in Review.

What's in the preview

Bring your own model

Connect OpenAI, Anthropic, DeepSeek, xAI, Kimi, Z.AI, MiniMax or Alibaba Cloud Model Studio with an API key. Gateways such as OpenRouter and SiliconFlow and local servers such as LM Studio work too, as does any custom OpenAI- or Anthropic-compatible endpoint. You can also reuse a subscription you already have: a Codex login, xAI SuperGrok or Grok Build.

Already use another coding agent? wuu can run Claude Code, Codex, Cursor, Devin, Grok, Hermes, OpenCode, Pi or Antigravity as the conversation's engine, with that agent's own login and tools. Settings → Subscriptions shows plan quotas and balances for every account in one place.

Keep the work in view

The workspace panel sits beside the conversation:

• Review shows the current Git diff, file by file.
• Files previews code, Markdown, images and documents.
• Terminal opens a shell in the project.
• Browser is the same page the agent drives. A floating preview follows its clicks, and clicking or typing on the page yourself pauses the task.

Select text in a file to add it to the conversation, leave a comment, ask about it in a side chat, or have the agent edit just that passage. Responses work the same way: quote a passage into your next message.

Steer without starting over

Press Enter while a task runs to steer it, or Command + Enter to queue the next message. Fork from any earlier message into the same folder or a new Git worktree. Start a conversation in its own worktree so experiments stay off your branch. Use /side for a quick question that stays out of the main thread, and search to find any past conversation.

Decide how far the agent can reach

Pick a permission mode for each conversation. Read-only is for investigation. Standard allows edits inside the workspace. Unconfined lifts the boundary when you really need it. In Read-only and Standard, the agent's commands run inside the macOS Seatbelt sandbox. Turn on Approve for me to have the model review shell commands and risky actions before they run.

When the work belongs somewhere else, run the agent's tools in Docker, over SSH, in Singularity/Apptainer, or in a Modal, Daytona or Vercel sandbox.

Make it yours

Plugins add tools and features: scheduled automation, persistent goals, task checklists, memory, subagents and more. Each one can be turned on or off. You can install a local plugin, add skills, connect MCP servers and run hooks. Extensions use the same public API as the bundled plugins. See Extend wuu.

Get started

The desktop preview runs on Apple silicon Macs. Download it from GitHub Releases, move wuu.app to /Applications, and open it.

The preview is unsigned. It has ad-hoc signatures only, with no Apple Developer ID or notarization. If macOS blocks it, follow the installation guide.

On first launch, connect a model service and add a project folder as a workspace. Start with a small task, then check the changes and test results in Review. The quick start walks through an example.

Command line

The desktop app includes its own core. To use wuu from a terminal or script as well, install the Go version listed in go.mod and build the CLI:

git clone https://github.com/blueberrycongee/wuu.git
cd wuu
make install
wuu init

Make sure Go's binary directory is on your PATH, and configure a model provider. Then run a task:

cd /path/to/your/project
wuu exec --permission-mode read_only "review this project and explain how to run its tests"

The wuu exec guide covers scripting, JSONL output and session controls.

Your files and data

wuu reads and changes local files and runs commands within the active permission mode. Prompts and relevant context go to the model provider you choose. Sessions and settings are stored under ~/.wuu by default. Read the security model before working with sensitive data or untrusted projects.

Contribute

See Contributing to work on wuu, or open an issue to report a problem. Report security vulnerabilities through SECURITY.md.

Acknowledgements

Thanks to @Rosemary1812 for contributing to Fusion's multi-agent collaboration, including Lead/Sidekick collaboration design, user experience, and behavioral testing (#489).

Licensed under MIT. Agent avatars use blobatar, also MIT-licensed. The screenshots above come from the real app working on synthetic example projects; landing/README.md explains how to regenerate them.